Document 11529683

advertisement
1
ANISTROPIC AUTOCORRELATION
IN HOUSE PRICES
by
Kevin Gillen
Thomas Thibodeau
Susan Wachter
Working Paper #383Anisotropic Autocorrelation in House Prices
Kevin Gillen
Real Estate Department
The Wharton School
University of Pennsylvania
314 Lauder-Fischer Hall
256 South 37th Street
Philadelphia, PA 19104.6330
Phone: 215.898.4818
Fax: 215.573.5261
Email: gillenk@wharton.upenn.edu
Thomas Thibodeau*
Professor of Real Estate
Cox School of Business
Southern Methodist University
Dallas, TX 75275-0333
Phone: 214.768.3613
Fax: 214.768.4099
Email: tthibode@mail.cox.smu.edu
Susan Wachter
2
Assistant Secretary Designee for
Policy Development and Research
US Department of Housing and Urban Development
451 Seventh Street SW
Suite 8100
Washington, DC 20410
Phone: 202.708.1600
Fax: 202.619.8000
*
Contact Author
October 2001
Abstract
This paper examines anisotropic spatial autocorrelation in single-family house prices and in
hedonic house price equation residuals using a spherical semivariogram and transactions data for
one county in the Philadelphia, Pennsylvania MSA. Isotropic semivariograms model spatial
relationships as a function of the distance separating properties in space. Anisotropic
semivariograms model spatial relationships as a function of both the distance and the direction
separating observations in space.
The goals of this paper are: (1) to determine whether there is spatial autocorrelation in the
hedonic residuals; and (2) to empirically examine the validity of the isotropy assumption.
We estimate the parameters of spherical semivariograms for house prices and for hedonic house
price equation residuals for 21 housing submarkets within Montgomery County, Pennsylvania.
These housing submarkets are constructed by dividing the entire county into 21 groupings of
economically similar adjacent census tracts. Census tracts are grouped according to 1990 census
tract median house prices and according to characteristics of the housing stock. We fit the
residuals of each submarket hedonic house price equation to both isotropic and anisotropic
spherical semivariograms. We find evidence of spatial autocorrelation in the hedonic residuals in
3
spite of a very elaborate hedonic specification. Additionally, we have determined that, in some
submarkets, the spatial autocorrelation in the hedonic residuals is anisotropic rather than
isotropic. The empirical results suggest that the spatial autocorrelation in single-family house
prices and in hedonic house price equation residuals is anisotropic in submarkets where most
residents commute to a regional or local Central Business District (CBD).
We acknowledge Professor Tony Smith, John Green of REALIST, and Paul Amos for their
assistance with this paper.
1. Introduction
House prices are spatially autocorrelated because properties in close proximity properties tend to
have similar structural characteristics (e.g. square feet of building/living area, dwelling age, and
design features). This is a natural consequence of the fact that spatially proximate properties
tend to be developed at about the same time. In addition, properties within the same
neighborhood share important neighborhood amenities (e.g. neighborhood properties have access
to the same public schools and are served by the same municipal police and fire departments).
Finally, house prices are likely to be spatially autocorrelated in neighborhoods where residents
follow similar commuting patterns.
House price models attempt to explain spatial and/or temporal variation in house prices. These
models are frequently used to mark residential property values to market. Hedonic models relate
house prices to characteristics of the lot, the structure, and the neighborhood (Gillingham [1975],
Goodman [1978], Thibodeau [1989, 1992, 1996], and many others). Repeat sales models
measure changes in house prices as the average rate of appreciation for properties that have sold
at least twice and have not undergone major structural changes between sales dates (Bailey,
Muth and Nourse [1963], Case and Shiller [1987, 1989]). Hybrid models combine hedonic and
repeat sales specifications to obtain more efficient parameter estimates (Case and Quigley
[1991], Quigley [1995], and Hill, Knight and Sirmans [1997]). Assessed value models estimate
price indices using information obtained from property tax assessment departments (Clapp and
Giaccotto [1992]).
4
The parameters of hedonic house price equations are typically estimated using Ordinary Least
Squares (OLS). This estimation procedure assumes the residuals are independently and
identically distributed with zero mean, a constant variance, and zero covariance. When the
residuals are spatially autocorrelated, the assumption of a zero covariance is violated and OLS
yields inefficient parameter estimates. More accurate parameter estimates can be obtained by
explicitly modeling the spatial autocorrelation. Modeling spatial relationships in hedonic house
price equations can also significantly improve market value prediction accuracy.
The residuals produced by house price models may be spatially autocorrelated for three reasons.
First, proximity externalities influence the market values of nearby properties in similar ways.
Thibodeau [1990] demonstrated that high-rise office buildings reduce the market value of nearby
homes by as much as 15%. Information on the determinants of proximity externalities in the
single-family market is difficult and costly to obtain. Second, other information on important
structural and neighborhood characteristics are not readily available and are excluded from
empirical house price specifications. For example, data on the quality of local public schools
and on area crime rates are difficult and costly to obtain, particularly on a large (e.g. national)
scale. In addition, information on neighborhood socioeconomic and demographic characteristics
are typically obtained from the U.S. Bureau of the Census. This information is collected only
once every ten years. When hedonic equations are used to model house prices, the residuals will
contain information on these unobserved housing characteristics. Third, even in ideal situations
where all housing characteristic information is available, it is difficult to select the "correct"
model specification. For example, it is difficult to model how public school quality gets
capitalized into the price of single-family properties. Model mis-specification may also
contribute to spatially autocorrelated hedonic house price equation residuals.
Several researchers have developed hedonic house price models that examine spatial
autocorrelation in house prices and in hedonic house price model residuals. Dubin [1988]
assumed the residual correlation between properties is a negative exponential function of the
distance between them and estimated hedonic parameters using a maximum likelihood procedure
suggested by Mardia and Marshall [1984]. Can [1992] incorporated spatially lagged values of
5
house prices in the hedonic specification. The absolute influence that nearby properties have on
value is determined using an exogenously specified weighting matrix. Pace and Gilley [1997]
model spatial dependence in house prices using a simultaneous autoregressive (SAR) model.
SAR models explain variation in house prices as a function of property characteristics and
spatially weighted hedonic residuals of comparable properties. They derive the SAR model by
combining the OLS and grid estimator and then compare the OLS, grid, and SAR estimates of
market value. Basu and Thibodeau [1998] examine spatial autocorrelation in Dallas house prices
using a semi-log hedonic house price equation and a spherical autocorrelation function with data
for over 5,000 transactions of homes sold between 1991:4 and 1993:1. Properties are geocoded
and assigned to separate housing submarkets within metropolitan Dallas. Hedonic and spherical
autocorrelation parameters are estimated separately for each submarket using estimated
generalized least squares (EGLS). They conclude that house prices are spatially autocorrelated
throughout metropolitan Dallas but that hedonic house price equation residuals are spatially
autocorrelated in about half of the submarkets examined.
Most research on spatial autocorrelation in house prices has assumed that the correlation
structure is isotropic--a function of only the distance between properties. The direction
separating properties is ignored. Spatial data is anisotropic when spatial autocorrelation is a
function of both the distance and the direction separating points in space. Anisotropic
semivariograms have been examined by Journel and Huijbregts [1978], Oden and Sokal [1986],
Isaaks and Srivastava [1989] and Simon [1997]. Simon [1997], for example, provides an exact
test for anisotropic spatial autocorrelation. The exact tests are obtained by projecting the two
dimensional spatial observations onto a single axis making an angle θ with the east-west axis.
Spatial autocorrelations are then replaced by the conventional product-moment correlation
coefficient.
House prices and hedonic house price equation residuals may exhibit anisotropic spatial
autocorrelation. Residential location theory suggests that housing consumers tradeoff housing
and commuting costs when selecting a residence. To reduce commuting costs, residential
properties are developed initially around major transportation arteries. This pattern of residential
development may result in stronger spatial autocorrelation in house prices (and in hedonic house
6
price equation residuals) along major transportation arteries, and in the direction of the Central
Business District (CBD).
This paper examines anisotropic spatial autocorrelation for Montgomery County single-family
transactions. We fit the parameters of a spherical autocorrelation function to the empirical (or
sample) semivariogram for two directions in each of 21 housing submarkets in suburban
Philadelphia. Our results suggest that the spatial autocorrelation in transaction prices and in
hedonic house price equation residuals is anisotropic for some housing submarkets.
2.
Specification
2.1
The Hedonic House Price Specification
We model the relationship between house prices and housing characteristics using a semilogarithmic functional form. This specification regresses the log of transaction prices on a linear
combination of (possibly transformed) housing characteristics. The semi-log functional form is
given by:
V = e Xβ + ε
(1)
where V is property value, X is a vector of (possibly transformed) housing characteristics, ß is a
vector of unknown hedonic coefficients, and ε is the residual. When the residual variance is
constant and the residuals are spatially uncorrelated, ordinary least squares (OLS) yields best,
linear, unbiased estimators of the parameters in the transformed equation:
Z = log V = Xß + ε,
(2)
where ε ~ N(0, σ2I) so that Z ~ N(Xß, σ2I). OLS yields estimated coefficients
b = (XTX)-1XTZ,
2.2
Modeling Spatial Autocorrelation
where b ~ N(ß, σ2(XTX)-1).
(3)
7
When the residuals are spatially autocorrelated, E { ε ε ' } = Ω, a matrix with non-zero offdiagonal elements. In this situation, ß can be estimated with the generalized least squares (GLS)
estimator B = ( X T Ω
-1
-1
-1
X ) X T Ω Z. The empirical challenge is to estimate the elements
of Ω.
Let si = ( ai, bi ) denote the location of property i (e.g. ai denotes the longitude and bi the latitude
for property i). Let ξ ( si ) denote the hedonic price equation residual for a property located at si.
If the stochastic process is weakly stationary, the covariogram for the distribution of residuals is
C ( si - sj ) = Cov { ξ ( si ), ξ ( sj ) } for all ( si, sj ). Note that C (0) is the (assumed constant)
variance for the residual distribution. The semivariogram of the process is:
γ (si - sj ) = 0.5 Var { ξ ( si ) - ξ ( sj ) } = C (0) - C ( si - sj ).
(4)
Let h denote the (Euclidean) distance separating locations si and sj. Clearly, γ (- h ) = γ ( h ) and
theoretically γ ( 0 ) = 0. Empirically however, γ ( h ) is sometimes discontinuous near the origin
and γ ( h ) → θ0 > 0, as h → 0. The discontinuity, θ0 , is labeled the nugget.
Observations may eventually become spatially uncorrelated as the distance between them
increases. When this happens, the semivariogram stops increasing beyond some threshold and
becomes constant. That is, γ ( h ) → C*, as h → ∞. This limiting value, C*, is called the sill of
the semivariogram. The range of the semivariogram is the value h0 such that γ ( h0 ) = C*. So
the range of a semivariogram is the distance beyond which observations are spatially
uncorrelated. Finally, a semivariogram is isotropic if γ (si - sj ) is a function of only the distance
between si and sj, ||si - sj ||, and not the direction separating si and sj. Spatial data is anisotropic
when spatial autocorrelation is a function of both the distance and the direction separating points
in space. Geometric anisotropy occurs when the range varies with direction but the sill is
constant. Zonal anisotropy occurs when the sill varies with direction but the range is constant
(Isaaks and Srivastava [1989]). Mixed models are models where both the sill and the range vary
with the direction separating spatial observations.
8
The empirical (or sample) semivariogram examines how the spatial autocorrelation between
observations changes as the distance between observations increases. The method of moments
estimator for an empirical semivariogram (Matheron, [1963]) is:
[
g (h) = ∑ ξ ( si )−ξ ( s j )
N (h)
]
2
2| N (h)|
(5)
where the average is taken over N(h) = {( si, sj ): si - sj = h} and | N(h) | is the distinct number of
pairs in N(h). For irregularly spaced data (e.g. single-family properties), N(h) is modified so that
N(h) = {(si, sj): si - sj ∈ T ( h )}, where T ( h ) is a tolerance region around h. Figure 1 plots the
points of the isotropic empirical (or sample) semivariogram for the log of transaction prices in
the Ambler submarket, an area in the north-central region of Montgomery County. A point in
the empirical semivariogram (e.g. a dot in Figure 1) is the difference between the variance and
covariance in (the log of ) house prices computed for properties within a given tolerance region.
The first point to the right of the vertical axis measures the spatial autocorrelation for properties
within +/- 200 meters of the separation distance h = 250 meters. The statistic is computed
without regard for the direction separating properties. The points for the empirical
semivariogram are computed for 40 values of h ranging from h=250 meters to h=6,500 meters.
The tolerance ranges or bins are constructed so that each bin has approximately the same number
of property pairs.
<Figure 1 About Here>
The next challenge is to fit a functional form to the points of the empirical semivariogram. Three
popular isotropic semivariograms used to empirically examine spatial relationships are the
spherical, exponential, and Gaussian semivariograms. (Cressie [1993] provides the functional
forms for these semivariograms. Additional references on modeling spatial autocorrelation
include Bailey and Gatrell [1995], Cliff and Ord [1973], Isaaks and Srivastava [1989], Ripley
[1981], and Dubin, Pace and Thibodeau [1998]). The spherical semivariogram model has a finite
range while the exponential and Gaussian semivariograms asymptotically approach a limiting
value. The functional form for the spherical model is given by Cressie [1993]:
9
h=0 


0 <|| h ||≤ θ 2 


|| h || ≥ θ 2 
0,

3
  || h || 
 || h ||  

 - 0.5
 ,
γ (h; θ ) = θ 0 + θ 1 1.5
  θ 2 
 θ 2  


θ 0 + θ 1 ,
(6)
The nugget for the spherical semivariogram is θ0, the sill is θ0 + θ1, and the range is θ2.
The parameters of the spherical semivariogram are estimated using nonlinear least squares. The
three parameters of the spherical semivariogram model can be fit to the empirical
semivariogram, g ( h ), by minimizing the nonlinear function:
K
S (θ ) = ∑ [ g ( h( k )) − γ ( h( k ),θ )]
2
(7)
k =1
with respect to the semivariogram parameters θ . The sequence h(1), ..., h(K) denotes the
separation distances for which the sample semivariogram g(h) are computed. Figure 1 also plots
a spherical semivariogram fitted to the points of the empirical semivariogram for the Ambler
submarket. The fitted spherical semivariogram is discontinuous at the origin (with an estimated
nugget = 0.0747) and increases with separation distance h. The fitted spherical semivariogram
becomes horizontal for h > 4.38 km, so house prices for properties separated by more than 4.38
km are spatially uncorrelated. If Ambler house prices were spatially uncorrelated, then the fitted
semivariogram would be horizontal at all separation distances.
To estimate the standard errors of the semivariogram parameter estimates, we assume the
semivariogram residuals are independently and identically distributed with mean zero and
variance σ2, then
K
σˆ 2 =
∑ (G
i =1
i
- γ i )2
K
G i = value of the empirical semivariogram at D i 
where : 

γ i = value of the spherical semivariogram at D i 
^
and the nonlinear least squares estimator θ
(8)
10
is approximately normally distributed with mean θ and covariance matrix (Maddala [1977])
^
^
σˆ 2 [F(θ ) T F(θ )]-1
(9)
F(θ) is an Nx3 gradient matrix created by evaluating the partial derivatives of (6) with respect to
θ0, θ1, and θ2 at each of the K categorical distances Di:
 ∂γ ∂γ ∂γ 
F(θ ) = 

 ∂θ 0 ∂θ 1 ∂θ 2 

= 1


3.
 3  D
i
 
 2  θ 2
 1  Di
 - 
 2  θ2



3



 3  D
θ 1 −  2i
 2  θ 2
 3  D 3i  
 +  4   for i = 1, 2, ... K
 2  θ 2  
(10)
The Data
This paper examines spatial autocorrelation in house prices and in hedonic house price equation
residuals using data for 21,562 transactions of single-family homes sold between 1995:Q1 and
1998:Q3 in Montgomery County, Pennsylvania. Montgomery County lies directly to the
northwest of Philadelphia. The county contains many homes that pre-date the American
Revolution. The eastern half of the county is a collection of various bedroom communities of
educated professionals. It includes portions of Philadelphia's famous "Main Line", a sequence of
affluent communities to the west of Philadelphia dating back to the 19th century. The region
includes several of the "Seven Sisters" schools (Bryn Mawr, Villanova, Swarthmore, etc.). The
western half of the county is predominately rural, containing several farming communities that
date to colonial times, including a strong Amish presence. Map 1 illustrates the location of
Montgomery County, PA.
<Map 1 About Here>
11
Realist, a private data vendor, provided 21,562 transactions of Montgomery County singlefamily properties. The typical Montgomery County single-family transaction had about 2,030
square feet of building area. The mean transaction price over the 1995-1998 period was
$179,823, or $88.61 per square foot. While a few homes were built prior to the American
Revolution, the average age for a Montgomery County property sold during the 1995-1998
period is 38.2 years.
3.1 The Submarkets
Twenty-one submarkets were defined by clustering contiguous census tracts with similar housing
characteristics. Within each submarket, we would ideally prefer a distribution of house prices
with a constant average unit price and a small variance, but the non-uniform spatial distribution
of properties constrained this possibility. That is, if we defined the geographic submarkets such
that each submarket could be classified as relatively "high-priced", "mid-priced", or "lowpriced", there would be several submarkets with very few observations in them. So, for
example, we were forced to occasionally include several low-priced tracts in an otherwise midpriced submarket. The result was 21 distinct geographic housing submarkets that generally
conformed with known municipal and geographic boundaries. Map 2 provides the submarket
boundaries for the 21 Montgomery County submarkets.
<Map 2 About Here>
Table 1 contains descriptive statistics for transactions in the 21 submarkets. The median submarket
transaction price ranged from $108,000 for Pottstown to $325,000 for the Main Line (West)
submarket. The mean per square foot price ranged from $73.33 for Pottstown to $126.80 for the
Main Line (West) submarket. The oldest properties are located in the eastern part of the county and
the average age declines for submarkets in the western portion of the county. For example, the
oldest properties are located in the Main Line (East) submarket (with an average age of 60 years)
while the newest stock is in the Salford submarket (with an average age of 20 years).
12
<Table 1 About Here>
Later in the paper we provide semivariogram parameter estimates for both isotropic and anisotropic
spherical semivariograms for each of the 21 submarkets. In addition, we provide detailed hedonic
results for three representative submarkets: the Main Line (East) submarket, the Norristown
submarket, and the Ambler submarket. A brief characterization of these three submarkets follows.
The Main Line is one of the oldest and wealthiest suburbs of the Philadelphia MSA. Named for a
still-active commuter train line that dates back to the 19th century, the Main Line is a contiguous
sequence of affluent suburbs that spans three counties, beginning at Philadelphia's westernmost city
line. To normalize the relative size of the submarkets, we divided the segment of the Main Line that
is within Montgomery County it into two distinct submarkets: Main Line (East) and Main Line
(West). The Main Line's housing stock can be characterized by global homogeneity, but local
heterogeneity: uniformly high-priced, but the value of any particular property can still differ from its
neighbor substantially (e.g. $600,000 v. $400,000). The 1996 estimated average median census
tract household income for the Main Line (East) submarket is $136,256 and 40.7% of the submarket
population has a college degree. The submarket is approximately 7.5 km by 9.0 km and has an area
of 27,400 m2. A typical property in the Main Line (East) submarket has 2,488 square feet of
building area and sold for $277,604, or $111.66 per square foot. This area has some of the oldest
homes in the Philadelphia area, with 25% of the transactions built prior to 1925. There are 1,399
transactions for the Main Line (East) submarket.
The Norristown submarket includes the edge city of Norristown and surrounding tracts.
Norristown is an old industrial mill town dating back to the American Revolution that was built
on the banks of the Schuylkill river for access to waterpower. The housing stock in the city
proper contains many traditionally working-class row homes and cottages in high-density urbanlike settings that are typical of neighborhoods in larger Northeastern cities. Although
Norristown's industrial base is no longer extant, and the city has become steadily enveloped in
Philadelphia's spreading suburban sprawl, Norristown persists as the area of Montgomery county
with the lowest median income and highest percentage of minority residents. The 1996 estimated
average median census tract household income for this submarket is $63,391 and 16.2% of the
13
population has a college degree. The housing stock of Norristown can be characterized by both
global and local homogeneity: uniformly low-priced neighborhoods of similar rowhomes and
small homes. The Norristown submarket is approximately 9.8 km by 9.1 km and has an area of
35,139 m2. The area residents are primarily blue collar. The average transaction price in
Norristown was $128,247, or $81.67 per square foot. Transactions had an average of 1,639
square feet of building area and were 46 years old at the time of sale. There are 997 Norristown
transactions.
The Ambler submarket is the stereotypical postwar middle-class bedroom community of the
American suburban landscape. It is also the housing submarket that is most similar in
characteristics to the overall Montgomery housing market. Geographically, Ambler is the largest
of the three submarkets profiled here, and unlike the other two submarkets, it contains no real
geographically defined center. It is best described as a collection of similar suburban housing
developments near to, but without any real relationship to, the township of Ambler. The 1996
estimated average median census tract household income for this submarket is $98,255, and
29.8% of the population have a college degree. The housing stock of Ambler is best
characterized as globally heterogeneous, but locally homogenous: clusters of architecturally
similar and like-priced homes, but some relative variance between clusters of development due
to the timing and nature of general suburban growth. The Ambler submarket is approximately
13.1 km by 11.7 km and has an area of 55,490 m2. The average transaction price in Ambler was
$232,301, or $94.92 per square foot. Transactions had an average of 2,446 square feet of
building area and were 28 years old at the time of sale. There are 1,773 transactions for Ambler.
4.
Empirical Specifications
4.1
The Empirical Hedonic House Price Specification
With a few exceptions, the empirical specification is standard. The house price specification
relates the log of transaction price to various structural (dwelling size, dwelling age, number of
stories, etc.) and location (distance to CBD) characteristics and includes dummy variables for
sales date.
14
The continuous variables in the model are common to most traditional hedonic model
specifications: log of building square footage, frontage, and log of number of stories. We also
take the ratio of building square footage to lot square footage to measure the pricing of aesthetic
proportionality of the property. As this variable becomes larger, it increasingly denotes a large
house on a small lot. Hence the negative sign on this variable's coefficient denotes a discounting
for the lack of yard space and/or privacy. Additionally, we also take the ratio of total number of
rooms to building square footage to measure the effects of average room size in a property. In
accordance with the Victorian aesthetic of the time, many homes in Montgomery County built
during the 19th century contain a very tight partitioning of the building into a sequence of many
rooms that would be considered awkwardly small by today's standards. The positive sign on its
coefficient indicates that, for a fixed house size, today’s consumers generally prefer fewer large
rooms to numerous small rooms.
A set of dummy variables measure the effects of categorical housing characteristics. For example, a
dummy variable is created for each unique value of qualitative variables, like exterior material and
type of heating fuel. Additionally, dummy variables are created for discrete variables, like number
of fireplaces or number of bathrooms to allow for nonlinear relationships between these structural
characteristics and house prices. As a general rule, the category with the greatest number of
observations associated with it serves as the omitted variable in the empirical specification. There
are eight possible categories for exterior material. The dummy variable for aluminum exterior is
omitted from the specification since it had the largest percentage of observations associated with it.
The empirical specification includes variables designed to measure the influence that housing
vintage has on house price. Exhibit 1 illustrates the distribution for year of construction.
Although the volume of construction is certainly increasing over time as the population of the
region increased, the historical pattern of housing construction exhibits several cycles of
development. Measuring from trough-to-trough, 6 distinct cycles of construction emerge, which
are labeled: pre-1865, 1866-1887, 1888-1918, 1919-1945, 1946-1975, and 1976-1998. Each
wave of new construction after 1865 is approximately 25 years in length, and all properties built
during a particular cycle are characterized by a common set of hedonic characteristics:
architectural style, exterior material, heating fuel, etc.
15
To capture the unique and nonlinear effects each cycle, we take the interaction of a property's
age with a dummy variable that equals "1" if the property was constructed during that cycle. We
do the same with age-squared and age-cubed. The general specification of the formula is:
AGE i , j = age of the ith house * (dummy = 1 if built during jth cycle, = 0 otherwise)
AGESQ i , j = age - squared of the ith house * (dummy = 1 if built during jth cycle, = 0
otherwise)
AGECUBE i , j = age - cubed of the ith house * (dummy = 1 if built during jth cycle, = 0
otherwise), for j = 1,..., 6; i = 1,..., 21,562.
<Exhibit 1 About Here>
The specification also includes variables that measure the influence that distance to three centers
of economic activity have on house price. The three centers of economic activity are the City of
Philadelphia, the Montgomery County CBD measured at the King of Prussia and the distance to
the submarket center of economic activity.
Finally, the specification includes dummy variables for sales quarter beginning with the first
quarter of 1995 and ending with the second quarter of 1998. The omitted category is for
properties sold at the end of the period (1998:3).
Although we impose the same specification on all submarkets, we did consider allowing the
specification to vary across submarkets to be better tailored to variations in consumer
preferences across these submarkets. But, we rejected this strategy since we reasoned that it
would make us vulnerable to the accusation that we possibly tweaked the specification in order
to induce anisotropy in the residuals. By imposing the same large and uniform specification on
all submarkets, we reduce the probability of there being any spatial effects present in the
residuals. Hence, any spatial autocorrelation in the residuals is not spurious. Reducing the
number of variables in the specification increases the likelihood of the residuals being correlated
with any omitted variables: surely an undesirable result.
16
The empirical hedonic house price specification is:
Ln (TRANSACTION PRICEi )= ß0
+ ß1*DWELLING AGE + ß2*AGESQ + ß3*AGECUBE
+ ß4*(DWELLING AGE*DEVELOPMENT CYCLE)
+ ß5*(DWELLING AGE*DEVELOPMENT CYCLE)2
+ ß6*(DWELLING AGE*DEVELOPMENT CYCLE)3
+ ß7*LN(BUILDING SQUARE FOOTAGE)
+ ß8*(BUILDING SF/LOT SQUARE FOOTAGE)
+ ß9*(TOTAL ROOMS/BUILDING SF)
+ ß10*LN(NUMBER OF STORIES)
+ ß11*TYPE OF EXTERIOR
+ ß12*TYPE OF SPACE HEATING SYSTEM
+ ß13*TYPE OF HEATING FUEL
+ ß14*TYPE OF BASEMENT
+ ß15*NUMBER OF FIREPLACES
+ ß16*NUMBER OF BATHROOMS
+ ß17*GARAGE CAPACITY
+ ß18*FRONTAGE
+ ß19*CORNER LOT
+ß20*SOURCE OF WATER SUPPLY
+ß21*SEWER SYSTEM
3
+ ∑ φ i * DISTANCE TO CBD i
i =1
T
+ ∑ δ j * SOLD j + ξ i ,
j =1
(10)
17
where
DISTANCE1i = Distance of the ith house to the Philadelphia CBD
DISTANCE 2i = Distance of the ith house to the the Montgomery County CBD *
DISTANCE 3i = Distance of the ith house to its submarket' s CBD
* King of Prussia
and SOLDj = 1 if property sold in quarter j and is zero otherwise; j = 1995:1, 1995:2, ....,
1998:2.
There are a number of discrete categorical variables in the specification which necessitated the
creation of a series of dummy variables. As a general rule, we always chose that value as the
omitted category for which the number of observations was most numerous. The following chart
gives the value of the omitted discrete variables used in the specification:
<Table 1a About Here>
The descriptive statistics for the housing characteristics included in the hedonic specification for
the Main Line (East), Norristown, and Ambler submarkets are provided in Table 2.
<Table 2 About Here>
5. Estimation Results
5.1 House Price Spherical Semivariograms
We examine directional spatial autocorrelation in Montgomery County house prices (and in
hedonic house price equation residuals) by fitting a spherical function to the empirical
semivariogram computed using properties separated by a given direction. Our analysis examines
18
directional autocorrelation for two directions: north-south and east-west. Since few properties
are located exactly north-south (or east-west) of one another, we compute the spatial
autocorrelation for properties within a tolerance range of the desired direction. The tolerance
region used here is
+/- 45%. Consequently, the points for the empirical semivariogram for
property pairs separated by a northerly direction are computed for properties located 90o +/- 45o.
Similarly, the points for the empirical semivariogram for property pairs separated in an easterly
direction are computed for properties located 0o +/- 45o.
The following table presents the empirically estimated thetas and (approximate) t-scores of the
spherical semivariogram for log(house price) in the Ambler submarket. The measurement of θ2,
the range, has been re-scaled to kilometers.
19
Estimated Semivariogram Parameters: Thetas and (t-scores)
Log (House Price)
Main Line (East) Submarket
Isotropic Spherical Semivariogram
θ0
0.099003
(t-score)
(12.801)
θ1
0.08001
(t-score)
θ2
(t-score)
(9.1672)
4.4115
(5.7187)
Anisotropic Spherical Semivariogram
Properties North-South
θ0
(t-score)
θ1
(t-score)
θ2
(t-score)
0.099888
(11.746)
1.4489
(0.0084357)
85.228
(0.0084211)
Properties East-West
θ0
0.10862
(t-score)
(12.637)
θ1
0.044479
(t-score)
θ2
(t-score)
(4.9164)
3.3459
(3.1912)
20
The empirical semivariogram results clearly indicate that the spatial autocorrelation of house
prices in this submarket is anisotropic. The values of the spherical semivariogram parameters
for the isotropic semivariogram are statistically different from zero, indicating that house values
are correlated up to a distance of approximately 4.4 km. However, the anisotropic
semivariograms differ in their results: although the estimated semivariogram parameters are
uniformly significant for the "east" semivariogram, only θ0 is significant for the "north"
semivariogram. This suggests that the Main Line (East) spatial autocorrelation in house prices
exists solely along the east-west axis of the submarket. Figure 2 plots the isotropic and
anisotropic semivariograms.
<Figure 2 about here>
First, note that the isotropic semivariogram lies exactly between the anisotropic semivariograms.
This would suggest that the isotropic measurement of spatial dependence is an average of the
anisotropic measurements. This isotropic semivariogram illustrates that (the log of ) the Main
Line house prices are spatially autocorrelated up to a range of about 4.4 kilometers. Beyond 4.4
kilometers, (the log of) the Main Line house prices are spatially uncorrelated. But, the
anisotropic semivariogram for "east" indicates that Main Line house prices are spatially
autocorrelated up to a distance of only 3.3 kilometers. At first glance, the non-converging
anisotropic semivariogram for "north" would suggest that the spatial stochastic process of house
prices is nonstationary, and the actual range of spatial autocorrelation is much greater than 4.4
kilometers. However, the insignificant t-scores for its parameters suggest just the opposite. In
actuality, there is no spatial autocorrelation in house prices in a north-south direction. This
spatial stochastic process is just a pure 'nugget' effect as house prices vary randomly from one
property to the next with no covariance. All spatial autocorrelation occurs along the east-west
axis.
Cities generally develop outwards from their centers. Consequently, we might a priori expect
that the direction of greatest autocorrelation would be in the direction of the city's CBD. This is
exactly true for this submarket: Philadelphia lies directly across the eastern edge of this
submarket's border.
21
5.2
Hedonic Parameters
The parameters of the hedonic equation and the spherical semivariogram are estimated separately
for the twenty-one submarkets. Table 3 provides the OLS regression statistics for the Main Line
(East), Norristown, and Ambler submarkets. (Results for the other submarkets are available
from the authors upon request).
<Table 3 About Here>
In general, dwelling size and age explain most of the variation in the log of transaction prices.
t-statistics for the logarithm of square feet of building area coefficients range from 15.7 for
Norristown to 23.9 for Ambler. (The estimated coefficient for the log of building area for the
Montgomery County hedonic has a t-statistic of 78.6.) Estimated coefficients for dwelling age
polynomials are jointly statistically significant for each submarket and document different
depreciation patterns. The omitted type of exterior is aluminum, so the estimated coefficients
measure differences from this type of exterior. The frame, masonry, stone, and stucco exteriors
command premiums in all submarkets. Properties with electric heat and heat pumps receive
discounts relative to properties with gas heat. A fireplace is worth between three and eight
percent of the value of the property. The estimated coefficients for the ratio of building square
feet to lot size are all negative and highly significant. Estimated coefficients for square feet of
garage space are statistically significant with the expected influence in each of the submarkets.
Corner lots are capitalized into all house prices as a discount. Additional rooms per square foot
of building area commands premiums in Norristown and Ambler, but not in the Main Line. The
distance of any given property to the Philadelphia CBD is strongly significant for the entire
county, but is generally insignificant within any given submarket. However, the distance to the
submarket's CBD is strongly significant for all submarkets, thus reflecting the importance of
relative global and local effects. Finally, between 1995:1 and 1998:3, house prices have
increased 5% for Montgomery County properties. Similarly, prices increased over 11% for the
Main Line properties and over 7% for Ambler properties, but house prices have been essentially
constant for Norristown homes.
22
5.3 Spherical Semivariograms for Hedonic Residuals
Table 4 provides the semivariogram parameter estimates for the residuals from the
hedonic house price equations for all 21 submarkets. The top half of the table lists
semivariogram parameters for the isotropic spherical semivariogram while the bottom half of the
table presents the spherical semivariogram parameters for two anisotropic spherical
semivariograms.
<Table 4 About Here>
Isotropic and anisotropic estimates of the nugget are statistically significant in each of the 21
submarkets. Isotropic estimates of θ1 and θ2 (the range) are statistically significant in 9
submarkets while anisotropic estimates of these parameters are significant in 11 submarkets. For
5 of these 11 submarkets, the estimated parameters for θ1 and θ2 are statistically in only one
direction. In Pottstown, New Hanover, Bryn Athyn, and the Main Line (West) submarket,
hedonic house price equation residuals are spatially correlated only for properties separated in a
north-south direction. For Cheltenham, hedonic house price equation residuals are spatially
correlated only for properties separated by an east-west direction.
The range of the anisotropic spatially autocorrelated residuals also varies for several submarkets
that exhibit both north-south and east-west spatial autocorrelation. For example, the range of
spatial autocorrelation in the Main Line (East) submarket for properties separated in a northsouth direction is over one kilometer further than the range for properties separated by an eastwest direction.
6. Conclusion
This paper examines anisotropic spatial autocorrelation in house prices and in hedonic house price
equation residuals for twenty-one housing submarkets in suburban Philadelphia. We find that both
house prices and hedonic house price equation residuals are spatially autocorrelated and the spatial
autocorrelation changes with the direction separating properties for some submarkets.
Since most real estate development tends to spread outwards from existing city centers along
major transportation arteries, it is reasonable to expect that the direction of greatest spatial
23
autocorrelation occurs in the direction of the CBD. Our empirical results are consistent with this
hypothesis. For those submarkets where anisotropy obtains, several are inner-ring suburban
bedroom communities that are close to Philadelphia. For two of the submarkets (Bryn Athyn,
Cheltenham), the direction of spatial autocorrelation is directly towards the Philadelphia City
Center. For a third submarket (Main Line West), the direction of anisotropy is towards a major
transportation artery connecting the Philadelphia CBD with the County CBD, both of which are
major employment centers for the residents of this submarket. For two other submarkets on the
rural western edge of the county (Pottstown, New Hanover), the direction of spatial
autocorrelation is greatest towards the Pottstown city center, which is the local CBD for this area
of the county.
Especially interesting is the characterization of local versus global anisotropic effects. While the
global pattern of development is outward from the Philadelphia CBD, local development also
spreads away from local CBDs such as the county and submarket CBDs. The direction
anisotropy is influenced by which CBD is the primary shopping and employment center for
submarket residents. For bedroom communities of Philadelphia commuters, anisotropy obtains
in the direction of the Philadelphia CBD. For submarkets where residents are equally likely to
commute to either the MSA CBD or county CBD, anisotropy obtains in the direction of the
transportation artery connecting these two CBDs. Finally, for the more rural and isolated
submarkets where the local CBD is the primary retail and employment district, anisotropy
obtains in the direction of this CBD, unaffected by the larger spread of development outward
from Philadelphia. Further investigation into the estimation of the local v. global 'layers' of
anisotropy is an area for further research.
It is our ultimate goal to evaluate the benefits of explicitly modeling anisotropic spatial
dependence in the residuals of a hedonic estimation. However, we leave this to a future paper.
The goal of this paper was first to identify if anisotropy obtains in the residuals, given a wellspecified hedonic estimation. We leave the evaluation of the benefits obtained from explicitly
modeling anisotropic spatial dependence in the residuals to a future paper. Future research will
also perhaps allow the hedonic specification to vary in order to maximally exploit the presence
of anisotropically autocorrelated residuals in the prediction of house values.
24
References
Anselin, Luc. 1978. Spatial Econometrics: Methods and Models Dordrecht: Kluwer
Academic.
Bailey, Trevor C. and Anthony C. Gatrell. 1995. Interactive Spatial Data Analysis Addison
Wesley Longman Limited. Essex, England.
Bailey, M., R. Muth and H. Nourse. 1963. "A Regression Method for Real Estate Price Index
Construction" Journal of the American Statistical Association 58 (304):933-942.
Basu, Sabyasachi and Thomas Thibodeau. 1998. “Analysis of Spatial Autocorrelation in House
Prices” Journal of Real Estate Finance and Economics 17 (1):61-85.
Can, Ayse. 1992. “Specification and Estimation of Hedonic Housing Price Models” Regional
Science and Urban Economics 22:453-477.
Case, Bradford and John Quigley. 1991. “The Dynamics of Real Estate Prices” Review of
Economics and Statistics 83:50-58.
Case, K. and R. Shiller. 1987. "Prices of Single Family Homes Since 1970: New Indexes for
Four Cities" National Bureau of Economic Research Working Paper No. 2393.
Cambridge, MA.
________. 1989. "The Efficiency of the Market for Single-Family Homes" American
Economic Review 79 (1):125-137.
Clapp, J. and C. Giaccotto. 1992. “Estimating Price Trends for Residential Property: A
Comparison of Repeat Sales and Assessed Value Methods” Journal of the American
Statistical Association 87:300-306.
Cliff, Andrew D. and J. K. Ord. 1973. Spatial Autocorrelation London: Pion Limited.
Cressie, Noel A. 1993. Statistics for Spatial Data New York: John Wiley and Sons.
Dubin, Robin A. 1988. “Estimation of Regression Coefficients in the Presence of Spatially
Autocorrelated Error Terms” Review of Economics and Statistics 70:466-474.
Dubin, Robin A., R. Kelley Pace and Thomas G. Thibodeau. 1998. “Spatial Autoregression
Techniques for Real Estate Data” The Journal of Real Estate Literature
Gillingham, Robert. 1975. "Place to Place Rent Comparisons" Annals of Economic and Social
Measurement 4(1):153-174.
25
Goodman, Allen C. 1978. "Hedonic Prices, Price Indices, and Housing Markets" Journal of
Urban Economics 5(4):471-484.
Hill, R. C., J.R. Knight and C.F. Sirmans. 1997. “Estimating Capital Asset Price Indexes” The
Review of Economics and Statistics 79 (2):226-233.
Isaaks, Edward H., and R. Mohan Srivastava. 1989. An Introduction to Applied Geostatistics.
New York: Oxford University Press.
Journel, A. G. and C. J. Huijbregts. 1978. Mining Geostatistics London: Academic Press.
Maddala, G. S. 1977. Econometrics. New York: McGraw Hill.
Mardia, K. V. and R. J. Marshall. 1984. "Maximum Likelihood Estimation Of Models for
Residual Covariance in Spatial Regression" Biometrika 71:135-146.
Matheron, G. 1963. “Principles of Geostatistics” Economic Geology 58:1246-1266.
Oden, N. L. and R. R. Sokal. 1986. “Directional Autocorrelation: An Extension of Spatial
Correlograms in Two Dimensions” Systematic Zoology 35:608-617.
Pace, R. Kelley, and O.W. Gilley. 1997. “Using the Spatial Configuration of the Data to
Improve Estimation” Journal of the Real Estate Finance and Economics 14 (3):333-340.
Quigley, John. 1995. “A Simple Hybrid Model for Estimating Real Estate Indexes” Journal of
Housing Economics 4(1):1-12.
Ripley, Brian. 1981. Spatial Statistics New York: John Wiley and Sons.
Simon, Gary. 1997. “An Angular Version of Spatial Correlations, with Exact Significance
Tests” Geographical Analysis 29 (3):267-278.
Thibodeau, Thomas G. 1989. "Housing Price Indexes from the 1974-83 SMSA Annual Housing
Surveys" Journal of the American Real Estate and Urban Association 17:100-117.
________. 1990. "Estimating the Effect of High Rise Office Buildings on Residential Property
Values" Land Economics 66 (4):402-408.
________. 1992. Residential Real Estate Prices from the 1974-1983 Standard Metropolitan
Statistical Area American Housing Survey Mount Pleasant, Michigan: Blackstone
Books: Studies in Urban and Resource Economics.
________. 1996. "House Price Indices from the 1984-1992 MSA American Housing Surveys"
Journal of Housing Research 6(3):439-481.
Download